GAAP

A macrocyclic peptide built from D-amino acids and N-methylated residues has no place in a twenty-letter alphabet, and a DNA duplex bound to a protein has no amino acids at all. That is the wall residue-level accuracy estimation runs into: AlphaFold 2 and RoseTTAFold emit confidence estimates that outperform dedicated methods, but both index by residue, so molecules outside the canonical set cannot be fed to them. GAAP discards the residue abstraction entirely and defines its graph over heavy atoms, typed by Rosetta chemistry rather than by sequence identity.

GAAP was developed by Naozumi Hiranuma in the Baker Lab at the University of Washington's Institute for Protein Design, and is described in Chapter 7 of his 2022 doctoral dissertation, which expands the acronym as "Generalized Atom Accuracy Predictor" in the text while titling the chapter "General-purpose accuracy estimator for non-protein molecules." It extends the lab's earlier protein-only accuracy estimators, DeepAccNet and Pluto, from residue-resolution convolutions to an SE(3)-transformer operating directly on atomic coordinates.

A single network's parameters are learned across several million chemical structures spanning four families: protein monomers, macrocyclic peptides, protein-DNA complexes and protein-ligand complexes. That one checkpoint is then applied to molecule classes held out of training, including protein-peptide complexes, generated for testing only. The dissertation remains the sole published account of the model, and no code or weights have been located in the lab's public repositories.

#Key Features

• Atom graph in place of a residue vocabulary: Nodes are non-hydrogen heavy atoms carrying one of 65 Rosetta atom types, with edges to the 16 nearest neighbours annotated by bondedness and bond order. Noncanonical and N-methylated residues need no special case.
• Two accuracy readouts at once: The network predicts a 50-bin distribution over full-atom l-DDT and, per graph edge, a categorical distribution of pairwise distance offsets from -5 Å to 5 Å in 0.5 Å bins, so each estimate carries its own confidence.
• Selectable scoring regions: Because predictions are made per atom pair, a user can average over any chosen subset of edges. In binder design, where the target is static, scoring only the interface edges isolates what is actually being designed.
• Screening throughput: The network produces roughly eight macrocycle predictions per second on CPU, fast enough to filter millions of designed backbones.

#Technical Details

The SE(3)-transformer trunk uses four layers, four attention heads, 32 channels and three degrees; its final layer emits 82 channels, of which 50 feed the l-DDT head and 32 feed the edge head, where two atoms' vectors are concatenated and passed through three feed-forward layers. Training combines 22,928 protein sequence clusters filtered at 40% identity with decoys from RosettaCM, trRosetta and RoseTTAFold; 97 experimentally resolved macrocycles from the Cambridge Structural Database and in-house IPD structures, at roughly 5,000 decoys each; roughly 8,000 PDBbind 2018 protein-ligand complexes re-docked with GALigandDock at 30 decoys per target, whittled down from an initial 18,000 by discarding ligands over 50 heavy atoms, covalently bound complexes and docking failures; and 335 protein-DNA structures clustered at 80% identity (the dissertation gives 333 in one sentence and 335 in the next), also 30 decoys each. Protein-DNA and macrocycle data are oversampled 40-fold and 5-fold to offset their scarcity, decoys are drawn uniformly across each target's l-DDT range rather than by natural frequency, and a single loss combining both cross-entropy terms runs for 120 epochs. On 15 held-out protein-DNA targets the model ranked the crystal structure among the top five decoys in every case, and did the same for Cartesian-minimised crystals, against 14/15 and 9/15 without uniform sampling. On unseen macrocycles the crystal structure is usually scored best with a conservation probability above 0.90, and GAAP beats the Rosetta energy function on the Boltzmann-score metric while matching it on Spearman correlation.

#Applications

The immediate use is filtering: ranking sampled or designed conformations before committing to synthesis. For macrocyclic peptide design, GAAP scores backbones — including poly-alanine backbones, whose predictions still track the l-DDT of the sidechain-packed structures — so it can filter earlier in the pipeline than sequence design. For DNA-binding protein design, interface-restricted scoring selects docked complexes that beat Rosetta's ddG on both RMSD and fraction of conserved contacts. Both suit large in-silico screens needing an atom-level quality score for chemistry protein-specific estimators cannot parse.

#Impact

GAAP marks the point where this line of accuracy estimation moved from residues to atoms, anticipating the all-atom biomolecular models that followed. Its demonstrated scope is narrower than its training mix: results are reported for macrocyclic peptides and protein-DNA complexes only, with no protein-ligand evaluation despite PDBbind's presence in training, and RNA is motivated in the introduction but never assessed. On macrocycle-protein complexes the network scores accurate decoys highly but does not reliably penalise inaccurate ones, limiting it to enrichment rather than absolute assessment. Because the work appeared only as a dissertation chapter, with no peer-reviewed paper, released code or checkpoint, its influence has been on the lab's own design workflows rather than on the wider community.